www.nature.

com/scientificreports

OPEN Monitoring tropical forest carbon

stocks and emissions using Planet
satellite data
Ovidiu Csillik   1*, Pramukta Kumar1, Joseph Mascaro2, Tara O’Shea2 & Gregory P. Asner1
Tropical forests are crucial for mitigating climate change, but many forests continue to be driven
from carbon sinks to sources through human activities. To support more sustainable forest uses, we
need to measure and monitor carbon stocks and emissions at high spatial and temporal resolution.
We developed the first large-scale very high-resolution map of aboveground carbon stocks and
emissions for the country of Peru by combining 6.7 million hectares of airborne LiDAR measurements
of top-of-canopy height with thousands of Planet Dove satellite images into a random forest machine
learning regression workflow, obtaining an R2 of 0.70 and RMSE of 25.38 Mg C ha−1 for the nationwide
estimation of aboveground carbon density (ACD). The diverse ecosystems of Peru harbor 6.928 Pg
C, of which only 2.9 Pg C are found in protected areas or their buffers. We found significant carbon
emissions between 2012 and 2017 in areas aggressively affected by oil palm and cacao plantations,
agricultural and urban expansions or illegal gold mining. Creating such a cost-effective and spatially
explicit indicators of aboveground carbon stocks and emissions for tropical countries will serve as a
transformative tool to quantify the climate change mitigation services that forests provide.

Mitigating the effects of climate change is a critical societal objective now and in the forthcoming decades.
Tropical countries contribute to carbon emissions mainly through deforestation and forest degradation, which
accounts for approximately 10% of the world’s annual total carbon emissions1. National and international initia-
tives such as REDD+2 are dedicated to reducing carbon emissions from deforestation and forest degradation. To
achieve this objective, each nation’s carbon emissions resulting from deforestation and forest degradation need to
be quantified and tracked over time3. At such large geographic scales, a precise, cost-effective and high-resolution
means to monitor changes in aboveground carbon stocks is needed.
Traditionally, forest carbon stocks have been estimated using field plot networks by correlating tree struc-
tural characteristics (diameter, height and wood density) to aboveground carbon density (ACD) using allometric
equations4. While this approach may be suitable for local areas, airborne LiDAR (light detection and ranging) has
proven useful in extending ACD mapping estimation outside of necessarily limited field plot inventory networks5.
LiDAR measurements provide detailed three-dimensional information of the forest canopy height and structure,
and it was shown that LiDAR- and field-based estimation of carbon stocks have an agreement of ~90% when cali-
brated at 1-ha spatial resolution5,6. At this resolution, the 10% difference is smaller than the error usually encoun-
tered in field-based estimation of ACD7. LiDAR can be used, therefore, to extend the mapping of carbon stocks to
areas larger than field-based estimates can provide, and has been used successfully for ACD or biomass mapping
in tropical forested regions of South America8, Central America6, Africa9–11, Asia12 or oceanic islands13. Although
LiDAR can extend the analysis to larger areas, it also reaches a geographic limit determined by costs and logistical
aspects associated with the use of aircraft14. To overcome this, LiDAR is often used with optical remote sensing
data of various spectral and spatial properties to scale from airborne to full-coverage, satellite-based areas8,15–17.
Combining satellite images and other geospatial datasets with airborne LiDAR has become a common
approach to map forest ACD across regions that lack LiDAR measurements18,19. Different spatial resolutions
of satellite images have been used, from low and medium to high and very high spatial resolution. Baccini and
Asner15 made use of Moderate Resolution Imaging Spectroradiometer (MODIS) images in combination with air-
borne LiDAR to generate pantropical ACD maps of greater details and improved accuracy. Asner et al.8 combined
Landsat-derived metrics with Shuttle Radar Topography Mission (SRTM) elevation variables to calibrate an ACD
model using airborne LiDAR data for the entire country of Peru. Using canopy texture metrics derived from high

1
Center for Global Discovery and Conservation Science, Arizona State University, Tempe, AZ, USA. 2Planet Labs Inc,
San Francisco, CA, USA. *email: [email protected]

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 1

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

resolution Geoeye-1 and Quickbird images, Bastin et al.20 mapped aboveground biomass (AGB) for very hetero-
geneous African forest types. Hojas Gascon et al.21 used RapidEye (5 m) optical images to reduce the efforts in col-
lecting national field data for estimating AGB over Tanzania. Hirata et al.22 developed an object-based approach
to map the AGB in tropical forests of Cambodia using a combination of airborne LiDAR with QuickBird images.
While medium and lower resolution satellite images have global coverage and are usually free of charge, the usage
of high and very-high resolution satellite images in estimating ACD is limited to smaller study areas, due to their
costs and availability21.
This spatial limitation of very-high resolution satellite image availability has recently changed via the rapid
emergence and availability of Planet data streams23. Planet Labs Inc. operates the largest fleet of Earth imaging
satellites, with around 180 “Dove” satellites currently in orbit and imaging the entire Earth, every day. Dove
imagery have been successfully used to map coral reefs and seagrass ecosystems24, benthic habitats25, agricul-
tural environments26–28, and digital elevation models have been successfully generated from multi-view Dove
imagery29. Imaging the Earth daily using four spectral bands (blue, green, red and near-infrared) at a resolution
of 3.7 m will contribute in overcoming an important issue when using satellite images in estimating ACD, namely
the cloud coverage of tropical regions. To our knowledge, Planet Dove images have not been used yet to map
aboveground carbon stocks for tropical forests at large extents.
Estimating forest properties, like ACD or tree canopy height, from optical images is accomplished using var-
ious machine learning regression models30–33. Of these, the Random Forest (RF) algorithm34 has proven to be
superior to traditional techniques for carbon mapping applications, such as regionally stratified sampling and
upscaling35. Part of this success is because RF is non-parametric, robust to a high number of input variables and
insensitive to data skew36. RF regression techniques have been intensively used to map carbon stocks8,35, bio-
mass21,31 or tree canopy height37,38, for a broad range of spatial resolutions of satellite images. When dealing with
high and very high resolution images, besides spectral reflectance, band ratios and indices derived from these,
common features used in an RF regression are related to image texture, which detects forest canopy structural
heterogeneity and ultimately predicts variations in ACD20. Two very popular textural measures used as a remotely
sensed vegetation structure feature are the grey-level co-occurrence matrix (GLCM) texture39–42 and Fourier
transform textural ordination (FOTO)20,43–45. Although textural features have been applied to various sensors
such as IKONOS-242, Cartosat-1a46, SPOT-541, QuickBird47,48, WorldView-249, or RapidEye21, it was not tested on
how it performs using Planet Dove images for large scale mapping of ACD.
Building on the previous work of Asner et al.8, who developed a country-wide ACD modeling framework
for Peru, our study aims to estimate ACD at an unprecedented level of detail for the entire country of Peru by
making use of Planet Dove spectral and textural features in combination with airborne LiDAR, integrated into
an efficient RF-machine learning regression framework. Furthermore, we compared our ACD results with those
from Asner et al.8 to calculate the carbon emissions and sinks for the period between 2012 and 2017 and show
examples of significant carbon changes through time led by oil palm and cacao plantations, urban expansions and
gold mining. Achieving high resolution ACD estimates at 1-ha resolution will greatly contribute to a rapid and
cost-effective monitoring system of carbon emissions for REDD + initiatives.

Results
Top-of-canopy height (TCH) estimation and uncertainties.  The TCH estimates yielded an R2 of 0.75
and a root mean square error (RMSE) of 3.90 m when compared with LiDAR-based TCH validation samples
(Fig. 1a). Grouping the RMSE results into 10 bins with a fitted function led to estimated RMSE of less than 4 m
for RF-estimated TCH of less than 7 m. TCH values of 7 to 21 m had an estimated RMSE in the form of negative
parabola (opens downwards) with values from 4 to 4.9 m with the peak around the value of 13 m for estimated
TCH (Fig. 1b). Transforming the RMSE values into percentage RMSE of estimated TCH depicted a decline in
uncertainty of RF-estimated TCH with increasing TCH values (Fig. 1c). This decline is desirable in the context of
ACD mapping, since a vast amount of carbon stocks are stored in trees with heights of 15 m or more. Estimated
TCH of 8 m or less had a percent RMSE of more than 50%. This error decreased up to 30% for TCH of 8 to 15 m.
Estimated TCH of 15–21 m had 30% to 20% RMSE, while TCH values higher than 25 m had less than 15% RMSE
of the estimated values. These values for the high trees were similar with the errors encountered during field
measurement of TCH.

ACD estimation and uncertainties.  Transforming the RF-estimated TCH into ACD resulted in a nation-
wide ACD model with R2 of 0.70 and RMSE of 25.38 Mg C ha−1 when compared against the 1.3 mil hectares of
LiDAR transformed into ACD (hereafter, LiDAR-derived ACD) (Fig. 1d). This was the first source of error in
estimating the uncertainty of our ACD model. Again, grouping the RMSE into 10 natural breaks bins and fitting
a polynomial function led to estimated RMSE values of less than 26 Mg C ha−1 for estimated ACD values of less
than 40 Mg C ha−1. RMSE values between 26 and 30 Mg C ha−1 characterized ACD values of 40 to 110 Mg C ha−1.
For values higher than 110 Mg C ha−1 the RMSE decreased while ACD increased, up to 22.5 Mg C ha−1 RMSE for
150 Mg C ha−1 estimated ACD (Fig. 1e). In terms of percent RMSE of ACD, there was a 50% or more RMSE for
estimated ACD values of 60 Mg C ha−1 or less, 50 to 30% RMSE for ACD of 60 to 100 Mg C ha−1 (Fig. 1f). ACD
values of 100 to 120 Mg C ha−1 had 30 to 20% RMSE, while estimated values higher than 120 Mg C ha−1 had less
than 20% RMSE of its values.
To better estimate uncertainty in our ACD map, we also considered the second type of error attributable to
calibrating the LiDAR-measured TCH to field-measured ACD (Eq. (1)). The mean error of LiDAR-based esti-
mates of ACD had previously been reported at 11.6%8. Combining the two sources of error (Eq. (2)) resulted in a
nationwide map of estimated relative uncertainty, expressed as a percentage of estimated ACD for every hectare
(Fig. 2).

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 2

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

Figure 1.  Density scatter plot depicting the relation between RF-estimated TCH and LiDAR-measured TCH,
using more than 1.3 mil 1-ha validation samples (a). Uncertainty of RF-estimated TCH expressed as root mean
squared error (RMSE, in m) of RF-estimated TCH, with a polynomial function fitted (b). Decline in uncertainty
(in %) of RF-estimated TCH with increasing TCH values, with a natural logarithm function fitted (c). Density
scatter plot depicting the relation between RF-estimated ACD and LiDAR-measured ACD, using more than
1.3 mil 1-ha validation samples (d). Uncertainty of RF-estimated ACD expressed as root mean squared error
(RMSE, in Mg C ha−1) of RF-estimated ACD, with a polynomial function fitted (e). Decline in uncertainty (in
%) of RF-estimated ACD with increasing ACD values, with a natural logarithm function fitted (f).

High-resolution map of ACD and regional statistics.  The high-resolution map of Peru’s ACD revealed
a wide range of ACD values, from <1 Mg C ha−1 in the western dry deserted areas to 150 Mg C ha−1 in the highest
biomass lowland Amazonian forest in northeastern part of the country (Fig. 2). The total aboveground carbon
stock estimated for Peru is 6.928 Pg with a country-scale uncertainty <1%, a value similar to the one obtained
by Asner et al.8 for Peru, of 6.922 Pg. The estimated ACD ranged between 0 and 152.3 Mg C ha−1, with a mean of
53.91 Mg C ha−1, and standard deviation of 49.07 Mg C ha−1. The diversity of ecosystems in Peru drove the spatial
arrangement of aboveground carbon stocks that is highly dependent on elevation, geological substrate, soil fertil-
ity, hydrological characteristics and climate8,50. Three legal jurisdictions in Peru store more than 78% (5.424 Pg)
of country’s total aboveground carbon stocks, with Loreto sheltering 3.613 Pg (52.1%), Ucayali 0.99 Pg (14.2%),
and Madre de Dios with 0.82 Pg (11.8%) (Fig. 3). The other 22 regions combined harbor 1.5 Pg C, with Western
Amazonian regions of San Martin (0.3 Pg), Amazonas (0.24 Pg) and Cusco (0.23 Pg) having more than 0.2 Pg C.
Mean ACD values reached a maximum of 96.7 Mg C ha−1 in Madre de Dios, 96.3 Mg C ha−1 in Loreto and
94.1 Mg C ha−1 in Ucayali (Fig. 3). The highest standard deviations in ACD were found in regions that transition
between high-biomass tropical forests to the eastern slopes of the Andes, like Pasco, Cusco, Junin, Huanuco,
Amazonas, San Martin, and Puno (Fig. 3).
From the total aboveground carbon stock of 6.928 Pg C estimated for Peru, 2.90 Pg C were found in five types
of protected areas, private and regional conservation areas, national protected areas, reserved areas, and buffer
zones (Fig. 4, Table 1). Of them, national protected areas with its buffer zones covered more than 32 mil hectares
with more than 2.5 Pg C. The highest mean ACD was for regional conservation areas (92.37 Mg C ha−1), which
had its three biggest conservation areas in high biomass forests of Loreto.

Carbon emissions between 2012 and 2017.  We compared our ACD estimates with results from Asner
et al.8 and developed a map of carbon stocks and emissions between 2012 and 2017 at 1-ha resolution. Using
high-resolution Planet Dove images facilitated a high precision view of specific landscape elements, such as small
rivers, as well as fine-scale disturbances and smooth transitions between vegetation types. Acknowledging the
uncertainties of the two ACD maps and the artifacts that might result in comparing the two, some areas across
Peru showed obvious signs of changed ACD that were identified by our analysis (Fig. 5). In Fig. 5, two examples

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 3

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

Figure 2.  The high-resolution ACD map of Peru at 1-ha resolution expressed in Mg C ha−1 using Planet Dove
satellite imagery (a). Estimated relative uncertainty expressed as a percentage of estimated ACD (b). The regions
of Peru are shown in black outlines with their associated names.

of natural sources of variation are represented by vegetation dynamics with insignificant changes between the two
time periods analyzed and river meandering that left a footprint of its hydrological evolution on the pattern of
carbon emissions (Fig. 5a–d). Human activities impacted the distribution of ACD and are represented mainly by
deforestation and forest disturbance along transportation routes, like in the case of Iquitos-Nauta road in Loreto
(Fig. 5e,f), newly emerged large areas of cocoa plantation disrupting intact forests (e.g. near Iquitos, Fig. 5g,h)
or oil palm plantations in areas that had already suffered major deforestation and land conversions (e.g. near
Pucallpa, Fig. 5i,j). One of the major threats to carbon stocks is gold mining in the Madre de Dios region, with
older mining areas expanding and new ones emerging at a rapid pace with irreversible impact over the environ-
ment (e.g. area between the cities of Boca Colorado and Puerto Maldonado, Fig. 5k,l).
Because the two maps were developed independently, we must be cautious in stating absolute values of carbon
emissions. While the examples of carbon changes from Fig. 5 are unquestionable, we combined the RMSE errors
of the two maps by computing the square root of the sum of the two, which resulted in a 41.84 Mg C ha−1 com-
bined error. Using this value as a threshold, our present map of carbon changes estimated 0.08 Pg C as a carbon
sink and 0.096 Pg C as carbon emissions between 2012 and 2017, while 0.02 Pg C fall between −41.81 and 41.84
Mg C ha−1 (Fig. 6). Regarding the area occupied by these differences, carbon emissions take 1.7 mil hectares,
carbon sink 1.5 mil hectares, while the rest of 125 mil hectares are attributable to uncertain differences (Fig. 6).

Discussion
We developed the first large-scale very high-resolution map of aboveground carbon stocks and change for the
country of Peru by combining 6.7 million hectares of airborne LiDAR measurements of top-of-canopy height
with thousands of Planet Dove satellite images into a random forest machine learning regression workflow,
obtaining an R2 of 0.70 and RMSE of 25.38 Mg C ha−1 for the nationwide estimation of ACD.
Random forest regression was chosen due to its proven accuracy and ability to deal with large datasets8,35.
Other studies have used an RF approach for estimating tree canopy height37, aboveground biomass21 or
aboveground carbon stocks8,35. For example, our RMSE is close to what Mascaro et al.35 found while mapping
tropical forest carbon in a 16 million hectare area in Western Amazon using RF with spatial context (RMSE of 26
Mg C ha−1) or similar to what Asner et al.8 found when estimating ACD at the country-scale of Peru (RMSE of
27.4 Mg C ha−1). We are aware that using local RF models fitted throughout Peru and then combining the results
will make our approach less transferable to other similar study areas. However, it was previously shown that RF
models trained using samples from all the study area can result in a universal model for canopy height estimation
with R2 higher than 0.6 and RMSE lower than 6 m38. Our trained model can be applied to different areas with
similar environmental conditions without the need for LiDAR data, or retrained for different forested ecosystems
based on LiDAR and satellite imagery.

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 4

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

Figure 3.  First level administrative subdivisions of Peru (regions) and their statistics of ACD, mean (a) and
standard deviation (b). The regions are ordered by the mean ACD (Mg C ha−1). The same regions of Peru
ordered by the total amount of aboveground carbon (in Pg C and % of total) and the cumulative graph are
shown in (c,d), respectively.

Creating a carbon estimation model using spectral and textural information from Planet Dove is changing
the way how we approach a cost-effective, automated, timely, and spatially-explicit indicator of carbon stocks and
emissions of tropical forests. Although we benefited from the high spatial and temporal resolution of the images,
the Planet Dove mission is still in its early stages. One major issue is related to the cross-sensor inconsisten-
cies26 that can lead to unwanted artifacts in the mosaic. These induced inconsistencies will propagate while using
sensitive variables like the GLCM textures. Using more robust features extracted from Dove mosaic for TCH
estimation will be one path for further investigations. We are also aware of the temporal differences between the
datasets used, with LiDAR from 2011–2013 and Dove images from 2017. Using RF with multiple decision trees
will neutralize the small number of LiDAR-sampled hectares that changed in this timeframe.
We estimated 6.928 Pg C stored as aboveground carbon stocks in Peru, a quantity similar to what Asner et
al.8 mapped for the same country. While only 2.90 Pg C are found in protected areas or their buffer areas, many
more high-carbon densities tropical forests are under threat by human activities, like oil palm plantation, gold
mining or fossil fuel oil extraction. To achieve a net neutral carbon balance, we need not only to limit gross
emissions of carbon, but also to transform more disturbed areas into carbon sinks. This is of critical importance
in the context of severe and frequent droughts the Amazon forests have experienced, which diminishes the role
of the Amazonian forests acting as carbon sinks17. For this, monitoring the carbon sequestrations and emissions
through time will help for better actions and we showed in this study a first-time high-resolution Peru-wide esti-
mates of carbon changes through time.

Conclusion
The role of tropical forests in mitigating the effects of climate change is important and needs to be better under-
stood. Here, we presented a large-scale mapping of aboveground carbon stocks at an unprecedented level of
detail by combining high resolution Planet Dove images and airborne LiDAR into a cost-effective and robust
random forest statistical approach. Furthermore, we showed a 1-ha resolution map of carbon emissions between

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 5

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

Figure 4.  The total aboveground carbon stocks (Mg C ha−1) found in five types of protected regions across
Peru.

Area Mean ACD STD ACD Total ACD

Protection type (mil ha) (Mg C ha−1) (Mg C ha−1) (Pg C)
Private conservation areas 0.38 19.45 22.17 0.007
Regional conservation areas 3.09 92.37 43.99 0.285
National protected areas 18.44 83.35 39.36 1.537
Buffer zones 13.98 73.04 43.23 1.021
Reserved areas 0.63 78.87 43.26 0.049

Table 1.  Statistics of estimated ACD for multiple types of protected areas in Peru, totaling 2.90 Pg C.

2012 and 2017, which can help improve understanding of land use changes and inform a sustainable pathway to
economic development for tropical countries. Creating cost-effective, automated and spatially explicit indica-
tors of aboveground carbon stocks for tropical countries has transformative potential for practically quantifying
the climate change mitigation services forests provide, including within MRV (Measurement, Reporting, and
Verification) systems for REDD + (Reducing emissions from deforestation and forest degradation) under the
UNFCCC (United Nations Framework Convention on Climate Change).

Data and Methods

Study area.  Our study area is the entire Republic of Peru, covering more than 128.5 million hectares. Forests
of Peru are very high in biodiversity, with high mountainous Andean regions extending from northwest to south-
east and tropical lowlands of the Amazonian Basin. These tropical forests are amongst the most biologically
diverse regions in Amazonia51, with high tree species richness, reaching more than 300 species with diame-
ter higher or equal with 10 cm in single hectares52. The highly diverse environmental and biological gradients,
together with rapid land use changes specific for a developing economy, strongly influence the carbon storage
throughout its ecosystems8.

Airborne LiDAR data.  The airborne LiDAR data was acquired during 2011 and 2013 flight campaigns using
the Global Airborne Observatory (GAO; formerly Carnegie Airborne Observatory)53 (Fig. 7). The GAO LiDAR
is a dual-laser scanning waveform system capable of firing 500,000 laser shots per second, with up to four discrete
returns per laser shot53. The aircraft was operated at an altitude averaging 2000 m above ground level with speeds
of up to 150 knots. An average-on-the-ground LiDAR points spacing of 4 shots per square meter was achieved,
with up to 8 shots per square meter in overlapping flight areas8. Thus, a 3D LiDAR point cloud with a resolution
of 1.12 m was achieved, covering not only forested ecosystem, but also other less representative ecosystems, like
grasslands, shrublands or savannas. The airborne LiDAR data were acquired on similar season as the PlanetScope
imagery, thus minimizing possible differences due to seasonality of the vegetation.
From the 3D LiDAR point cloud, a digital terrain model (DTM) and a digital surface model (DSM) were
constructed from the last and first returns, respectively. Subtracting DTM from DSM generated a top-of-canopy

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 6

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

Figure 5.  Examples of changes over time in ACD (Mg C ha−1), measured as the difference between our ACD
estimates (2017) and the results from Asner et al.8 from 2012. Intact forests showed insignificant changes in
ACD in southeastern Peru (a,b), while other sources of natural variation, like river meandering, created a
buffer of carbon emission zones along its path (c,d). Human-dominated sources of carbon emissions were
represented by deforestation along the Iquitos-Nauta road in Loreto (e,f), aggressive cocoa plantation expansion
disturbing a high biomass forest near Iquitos (g,h), deforestation, forest degradation and two new large areas of
oil palm plantation near the city of Pucallpa, Ucayali (i,j), and expansion of gold mining activities between Boca
Colorado and Puerto Maldonado, in Madre de Dios (k,l).

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 7

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

Figure 6.  Statistics of carbon sink and emissions between 2012 and 2017, in terms of total amount of changed
aboveground carbon stocks (a) and area occupied by the differences between the two maps (b). Vertical lines
depict the combined RMSE errors of the two maps (41.81 Mg C ha−1).

Figure 7.  Flight paths and airborne LiDAR data acquired by the Global Airborne Observatory (white paths)
overlapping a false color composite of Planet Dove mosaic (a). Two zoomed-in subsets of LiDAR (b) and Dove
(c) are shown for a region that underwent forest disturbances through deforestation and gold mining. Both
LiDAR and Dove mosaic are shown here at 1-ha resolution, as used in the random forest (RF) regression.

height (TCH) model at a spatial resolution of 1.1 m, covering 6,677,177 hectares throughout Peru (Fig. 7). This
simple LiDAR metric, TCH, was shown to be effective in mapping ACD at 1-ha resolution5.

Planet Dove satellite data.  Planet Dove is a low-Earth orbital constellation comprised of approximately
180 CubeSat 3U form factor (10 cm by 10 cm by 30 cm) satellites operating in sun-synchronous orbit (475 km
altitude). Dove satellites acquire images using four spectral bands, blue (455–515 nm), green (500–590 nm), red
(590–670 nm) and near infrared (780–860 nm), at a ground sample distance at nadir of approximately 3–4 m and

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 8

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

a positional accuracy of less than 10 m root mean square error (RMSE)54. To generate a seamless analysis-ready
mosaic of Dove images for Peru, we combined analytic ortho scene products that previously had top of atmos-
phere radiance correction, surface reflectance atmospheric correction, have been normalized, resampled to
3.125 m and orthorectified using GCPs and fine DEMs to less than 10 m RMSE positional accuracy. We gener-
ated the mosaic using 73,466 Dove scenes from the dry season of 2017 (July 1 - September 30), which offered
cloud-free pixels. The final cloud coverage of the mosaic was 0.8%, with isolated clouds over the Andes. (Fig. 7).
The normalized mosaic was created by applying a transformation to Dove surface reflectance (SR) data based
on a linear fit of each scene’s data to co-registered Landsat data from a similar season. The transform is con-
strained to preserve values with a reflectance of 1.0, and to prefer darkening the scene to brightening it. The latter
is because the MODIS AOT (Aerosol Optical Thickness) maps used for SR more often lead to underestimates of
the atmospheric component than overestimates. As a result, the Dove SR scenes are frequently too bright, particu-
larly in tropical areas. As a final step (optional, but recommended), a seamline removal algorithm was applied to
make a long-wavelength adjustment to intensity near scene boundaries so that the values in adjacent scenes are
similar, with values near a scene boundary changing more than values away from a scene boundary (i.e. gradient
reconstruction). Note that the seamline removal is not blurring or feathering and does not affect the spatial reso-
lution of the data. It does, however, shift the absolute values.

Dove GLCM texture.  Image texture is an important characteristic of every image that can help in identifying
different objects or regions within an image by measuring the spatial arrangements of image tone intensities39.
Image texture metrics have been widely used to measure vegetation structure characteristics40,42 and, furthermore,
to estimate forest biomass and carbon stocks55,56. We used the gray level co-occurrence matrix (GLCM) texture39,
which are one of the most popular texture metrics for remote sensing applications48. GLCM are second-order
texture measures that compute how often pairs of pixels with similar brightness values (gray tones) appear in an
image at a given spatial relationship. We computed eight GLCM texture measures, namely the mean, variance,
homogeneity, contrast, dissimilarity, entropy, second moment and correlation using a window size of 3 × 3 pixels
and a shift of 1 pixel, for every pixel in the image. We chose a small window size to capture the hard-to-detect
detailed changes in the dense canopy structure of the tropical forests. We used 32 levels of gray and derived the
texture in a single direction, of 135°. We used the Dove green band to compute the GLCM textures. All of these
settings were decided after running tests on smaller areas to find the best solution in terms of accuracy and speed
of computation.

Random forest regression.  Random Forest (RF) is an ensemble technique used for both classification and
regression and have become widely popular for remote sensing applications due to the accuracy of its results34,57,58.
RF uses multiple decision trees that are built independently using subsets of training samples, which are drawn
through replacement (bagging). The decision trees are then combined into a ‘forest’ to predict the final output,
performing better than individual decision trees. Unlike linear models, RF can capture complex non-linear rela-
tionships between predictors and the target variable.
We used 12 predictors in the RF regression to estimate TCH, namely green, red and near-infrared bands,
the eight GLCM textures and SRTM elevation (Fig. 8, Supplementary Fig. 1). We did not include the blue band
since it’s the most sensitive to atmosphere and might introduce unwanted artifacts in the analysis. After being
co-aligned and stacked, the 1-ha resolution layers covering the entire Peru were tiled mostly into 300 × 300 km
tiles, with larger tiles towards the land borders or Pacific coast of Peru, where the LiDAR coverage was sparser.
These tiles have an overlap of 50% (150 km) on each side, meaning that pixels further away from the borders were
predicted 4 times by the RF regression. To account for the subjectivity of tiles location when creating them, each
tile was shifted with 25% (75 km) towards West, East, South, and North.
The 6,677,177 ha of LiDAR data were split into 80% (5,341,742 ha) for RF regression training and validation
and 20% (1,335,435 ha) were kept for the final validation of the country-wide TCH estimation and were not used
in any of the RF workflow (Fig. 8). From here onwards, we call a local RF model a model that was applied for one
tile of approximate 300 × 300 km. For each local RF model, we used 40,000 samples of 1-ha LiDAR TCH values
for training the model, while the remaining samples were used to validate the local model. Two main parameters
need to be set up in an RF regression approach, namely the number of trees to grow the forest (Ntree) and the
number of variables randomly sampled as candidates for each split (mtry)59. For the latter, we used 4 variables to
be considered at each split (p/3, where p is the number of predictors), while growing 250 trees. This number of
trees were decided after running multiple country-wide tests that showed no significant improvement in error
rate by increasing the Ntree beyond 250 trees. This lower value also ensured a faster computation for the RF
regression.
After training and validating more than 200 local RF models (Supplementary Fig. 2), five country-wide mod-
els of TCH estimates were produced for each case of tiles positions (unshifted, shifted W, E, S, N) (Supplementary
Fig. 3). To minimize the discontinuities between the tiles, a mosaicking procedure based on blending the pixel
values of the overlapping areas was used. The blend values are the result of a weighting procedure that consider
the distance from the pixel to the edge of the tile, within the overlapping area. The five TCH models were then
combined into a single TCH mosaic by averaging the overlapping pixel values. To ensure the robustness of our RF
approach, the workflow was repeated 10 times and these models too were averaged to obtain the final Peru-wide
TCH estimation (Fig. 8).

Estimating ACD and associated uncertainties.  Converting the estimated TCH to ACD was done using
the calibration equation proposed by Asner et al.8 (Eq. (1)). This equation was developed using a permanent
inventory plot network located in diverse ecosystems from which the ACD was derived using allometric equa-
tions. The calibration between LiDAR-derived TCH and field-measured ACD resulted in a mean error of 11.6%,

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 9

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

Figure 8.  Methodological workflow for aboveground carbon density (ACD) estimation.

in conditions of extreme heterogeneity of biological and land use diversity in the plot network8. In the end, our
Peru-wide maps of TCH and ACD were validated against the remaining 20% validation samples, from which
performance statistics were extracted.

ACD = 0.8245 × TCH1.573 (1)

We computed uncertainties for TCH and ACD estimation independently. In case of TCH, we grouped the
RMSE results into 10 bins using natural breaks method and fitted a function to obtain how the errors are changing
in relation to the estimated TCH, both in meters and percentage of the estimated TCH. In the case of ACD, we
used the same method to fit a function to model the estimation error in terms of both absolute (Mg C ha−1) and
percentage values of estimated ACD. For ACD, we combined this error with the second type of error that resulted
after calibrating the LiDAR-measured TCH to field-measured ACD by Asner et al.8 by computing the square root
of the sum of the two squared errors (Eq. (2)).

uc (ACD) = u(ACDfield)2 + u(ACDRFestimated)2 (2)

where u c (ACD) is the combined uncertainty for ACD, u(ACD field ) is the uncertainty of calibrating
LiDAR-measured TCH to field-measured ACD, and u(ACDRFestimated) is the uncertainty of our estimated ACD
using the RF approach.

Data availability
The data that support the findings of this study are available from Planet Labs Inc. and Global Airborne
Observatory (GAO), but restrictions apply to the availability of these data, which were used under license for the
current study, and so are not publicly available. GAO data is available upon request from G.P.A.

Received: 2 April 2019; Accepted: 13 November 2019;

Published: xx xx xxxx

References
1. Baccini, A. et al. Estimated carbon dioxide emissions from tropical deforestation improved by carbon-density maps. Nat. Clim.
Chang. 2, 182–185 (2012).
2. Corbera, E. & Schroeder, H. Governing and implementing REDD+. Environ. Sci. Policy 14, 89–99 (2011).
3. Gibbs, H. K., Brown, S., Niles, J. O. & Foley, J. A. Monitoring and estimating tropical forest carbon stocks: Making REDD a reality.
Environ. Res. Lett. 2 (2007).
4. Chave, J. et al. Tree allometry and improved estimation of carbon stocks and balance in tropical forests. Oecologia 145, 87–99 (2005).
5. Asner, G. P. & Mascaro, J. Mapping tropical forest carbon: Calibrating plot estimates to a simple LiDAR metric. Remote Sens.
Environ. 140, 614–624 (2014).

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 10

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

6. Mascaro, J., Detto, M., Asner, G. P. & Muller-Landau, H. C. Evaluating uncertainty in mapping forest carbon with airborne LiDAR.
Remote Sens. Environ. 115, 3770–3774 (2011).
7. Chave, J. et al. Error propagation and scaling for tropical forest biomass estimates. Philos. Trans. R. Soc. Lond. B Biol. Sci. 359,
409–420 (2004).
8. Asner, G. P. et al. Targeted carbon conservation at national scales with high-resolution monitoring. Proceedings of the National
Academy of Sciences 111, E5016–E5022 (2014).
9. Vaglio Laurin, G. et al. Above ground biomass estimation in an African tropical forest with lidar and hyperspectral data. ISPRS J.
Photogramm. Remote Sens. 89, 49–58 (2014).
10. Asner, G. P. et al. Human and environmental controls over aboveground carbon storage in Madagascar. Carbon Balance Manag. 7, 2
(2012).
11. Bouvet, A. et al. An above-ground biomass map of African savannahs and woodlands at 25m resolution derived from ALOS
PALSAR. Remote Sens. Environ. 206, 156–173 (2018).
12. Jucker, T. et al. Estimating aboveground carbon density and its uncertainty in Borneo’ s structurally complex tropical forests using
airborne laser scanning, 1–29 (2018).
13. Hughes, R. F. et al. Estimating aboveground carbon density across forest landscapes of Hawaii: Combining FIA plot-derived
estimates and airborne LiDAR. For. Ecol. Manage. 424, 323–337 (2018).
14. Asner, G. P. et al. High-resolution mapping of forest carbon stocks in the Colombian Amazon. Biogeosciences 9, 2683–2696 (2012).
15. Baccini, A. & Asner, G. P. Improving pantropical forest carbon maps with airborne LiDAR sampling. Carbon Management 4,
591–600 (2013).
16. Saatchi, S. S. et al. Benchmark map of forest carbon stocks in tropical regions across three continents. Proc. Natl. Acad. Sci. USA 108,
9899–9904 (2011).
17. Yang, Y. et al. Post-drought decline of the Amazon carbon sink. Nat. Commun. 9 (2018).
18. Asner, G. P. et al. Mapped aboveground carbon stocks to advance forest conservation and recovery in Malaysian Borneo. Biol.
Conserv. 217, 289–310 (2018).
19. Asner, G. P. Tropical forest carbon assessment: Integrating satellite and airborne mapping approaches. Environ. Res. Lett. 4 (2009).
20. Bastin, J. F. et al. Aboveground biomass mapping of African forest mosaics using canopy texture analysis: Toward a regional
approach. Ecol. Appl. 24, 1984–2001 (2014).
21. Hojas Gascón, L., Ceccherini, G., García Haro, F. J., Avitabile, V. & Eva, H. The Potential of High Resolution (5 m) RapidEye Optical
Data to Estimate Above Ground Biomass at the National Level over Tanzania. For. Trees Livelihoods 10, 107 (2019).
22. Hirata, Y. et al. Object-based mapping of aboveground biomass in tropical forests using LiDAR and very-high-spatial-resolution
satellite data. Remote Sensing 10 (2018).
23. Planet Team. Planet Application Program Interface: In Space for Life on Earth (2017).
24. Asner, G. P., Martin, R. E. & Mascaro, J. Coral reef atoll assessment in the South China Sea using Planet Dove satellites. Remote
Sensing in Ecology and Conservation 3, 57–65 (2017).
25. Wicaksono, P. & Lazuardi, W. Assessment of PlanetScope images for benthic habitat and seagrass species mapping in a complex
optically shallow water environment. Int. J. Remote Sens. 39, 5739–5765 (2018).
26. Houborg, R. & Mccabe, M. F. High-Resolution NDVI from Planet’ s Constellation of Earth Observing Nano-Satellites: A New Data
Source for Precision Agriculture, https://doi.org/10.3390/rs8090768 (2016).
27. Houborg, R. & McCabe, M. F. A Cubesat enabled Spatio-Temporal Enhancement Method (CESTEM) utilizing Planet, Landsat and
MODIS data. Remote Sens. Environ. 209, 211–226 (2018).
28. Shi, Y. et al. Partial Least Square Discriminant Analysis Based on Normalized Two-Stage Vegetation Indices for Mapping Damage
from Rice Diseases Using PlanetScope Datasets. Sensors 18, (2018).
29. Ghuffar, S. DEM Generation from Multi Satellite PlanetScope Imagery. Remote Sensing 10, 1462 (2018).
30. De’ath, G. & Fabricius, K. E. {Classification} {and} {Regression} {Trees:} a {Powerful} {Yet} {Simple} {Technique} {for} {Ecological}
{Data} {Analysis}. Ecology 81, 3178–3192 (2000).
31. Gleason, C. J. & Im, J. Forest biomass estimation from airborne LiDAR data using machine learning approaches. Remote Sens.
Environ. 125, 80–91 (2012).
32. Ali, I., Greifeneder, F., Stamenkovic, J., Neumann, M. & Notarnicola, C. Review of machine learning approaches for biomass and soil
moisture retrievals from remote sensing data. Remote Sensing 7, 16398–16421 (2015).
33. Prasad, A. M., Iverson, L. R. & Liaw, A. Newer classification and regression tree techniques: Bagging and random forests for
ecological prediction. Ecosystems 9, 181–199 (2006).
34. Breiman, L. Random Forests. Mach. Learn. 45, 5–32 (2001).
35. Mascaro, J. et al. A tale of two ‘Forests’: Random Forest machine learning aids tropical Forest carbon mapping. PLoS One 9, 12–16
(2014).
36. Evans, J. S., Murphy, M. A., Holden, Z. A. & Cushman, S. A. Modeling Species Distribution and Change Using Random Forest. in
Predictive Species and Habitat Modeling in Landscape Ecology: Concepts and Applications (eds. Drew, C. A., Wiersma, Y. F. &
Huettmann, F.) 139–159 (Springer New York, 2011).
37. Staben, G., Lucieer, A. & Scarth, P. Modelling LiDAR derived tree canopy height from Landsat TM, ETM+ and OLI satellite
imagery—A machine learning approach. Int. J. Appl. Earth Obs. Geoinf. 73, 666–681 (2018).
38. Jin, S. et al. The transferability of Random Forest in canopy height estimation from multi-source remote sensing data. Remote
Sensing 10, 1–21 (2018).
39. Haralick, R. M., Shanmugam, K. & Dinstein, I. Textural Features for Image Classification. IEEE Trans. Syst. Man Cybern. SMC-3,
610–621 (1973).
40. Wood, E. M., Pidgeon, A. M., Radeloff, V. C. & Keuler, N. S. Image texture as a remotely sensed measure of vegetation structure.
Remote Sens. Environ. 121, 516–526 (2012).
41. Meng, J. et al. Mapping Forest Health Using Spectral and Textural Information Extracted from SPOT-5 Satellite Images. Remote
Sensing 8, 719 (2016).
42. Kayitakire, F., Hamel, C. & Defourny, P. Retrieving forest structure variables based on image texture analysis and IKONOS-2
imagery. Remote Sens. Environ. 102, 390–401 (2006).
43. Couteron, P., Pelissier, R., Nicolini, E. A. & Paget, D. Predicting tropical forest stand structure parameters from Fourier transform of
very high-resolution remotely sensed canopy images. J. Appl. Ecol. 42, 1121–1128 (2005).
44. Ploton, P. et al. Assessing aboveground tropical forest biomass using Google Earth canopy images. Ecol. Appl. 22, 993–1003 (2012).
45. Barbier, N., Couteron, P., Gastelly-Etchegorry, J.-P. & Proisy, C. Linking canopy images to forest structural parameters: potential of
a modeling framework. Ann. For. Sci. 69, 305–311 (2012).
46. Pargal, S. et al. Inverting Aboveground Biomass–Canopy Texture Relationships in a Landscape of Forest Mosaic in the Western
Ghats of India Using Very High Resolution Cartosat Imagery. Remote Sensing 9 (2017).
47. Petrou, Z. I., Tarantino, C., Adamo, M., Blonda, P. & Petrou, M. Estimation of Vegetation Height Through Satellite Image Texture
Analysis. ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences XXXIX-B8,
321–326 (2012).
48. Zhou, J. et al. The Effects of GLCM parameters on LAI estimation using texture values from Quickbird Satellite Imagery. Sci. Rep. 7,
7366 (2017).

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 11

Content courtesy of Springer Nature, terms of use apply. Rights reserved
www.nature.com/scientificreports/ www.nature.com/scientificreports

49. Ozdemir, I. & Karnieli, A. Predicting forest structural parameters using the image texture derived from WorldView-2 multispectral
imagery in a dryland forest, Israel. Int. J. Appl. Earth Obs. Geoinf. 13, 701–710 (2011).
50. Asner, G. P. et al. High-resolution forest carbon stocks and emissions in the Amazon. Proceedings of the National Academy of Sciences
107, 16738–16742 (2010).
51. ter Steege, H. et al. Continental-scale patterns of canopy tree composition and function across Amazonia. Nature 443, 444–447
(2006).
52. Gentry, A. H. Tree species richness of upper Amazonian forests. Proc. Natl. Acad. Sci. USA 85, 156–159 (1988).
53. Asner, G. P. et al. Carnegie Airborne Observatory-2: Increasing science data dimensionality via high-fidelity multi-sensor fusion.
Remote Sens. Environ. 124, 454–465 (2012).
54. Planet Team. Planet Imagery Product Specifications (2018).
55. Eckert, S. Improved Forest Biomass and Carbon Estimations Using Texture Measures from WorldView-2 Satellite Data. Remote
Sensing 4, 810–829 (2012).
56. Cutler, M. E. J., Boyd, D. S., Foody, G. M. & Vetrivel, A. Estimating tropical forest biomass with a combination of SAR image texture
and Landsat TM data: An assessment of predictions between regions. ISPRS J. Photogramm. Remote Sens. 70, 66–77 (2012).
57. Belgiu, M. & Dragut, L. Random forest in remote sensing: A review of applications and future directions. ISPRS J. Photogramm.
Remote Sens. 114, 24–31 (2016).
58. Cutler, D. R. et al. Random Forests for Classification in Ecology. Ecology 88, 2783–2792 (2007).
59. Liaw, A. & Wiener, M. & Others. Classification and regression by randomForest. R news 2, 18–22 (2002).

Acknowledgements
This study was supported by a grant from the Erol Foundation. We thank J. Kington and D. Knapp for assistance
with the Planet Dove mosaics and LiDAR data, respectively. The GAO is made possible by grants and donations
to G. P. Asner from the Avatar Alliance Foundation, Margaret A. Cargill Foundation, David and Lucile Packard
Foundation, Gordon and Betty Moore Foundation, Grantham Foundation for the Protection of the Environment,
W. M. Keck Foundation, John D. and Catherine T. MacArthur Foundation, Andrew Mellon Foundation, Mary
Anne Nyburg Baker, G. Leonard Baker Jr, and William R. Hearst III.

Author contributions
O.C., P.K., J.M., T.O. and G.P.A. contributed to research design. O.C. conducted the analysis. O.C., P.K., J.M., T.O.
and G.P.A. discussed the results and their implications. O.C. and G.P.A. wrote the manuscript and P.K., J.M. and
T.O. reviewed and edited the manuscript.

Competing interests
The authors declare no competing interests.

Additional information
Supplementary information is available for this paper at https://doi.org/10.1038/s41598-019-54386-6.
Correspondence and requests for materials should be addressed to O.C.
Reprints and permissions information is available at www.nature.com/reprints.
Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Cre-
ative Commons license, and indicate if changes were made. The images or other third party material in this
article are included in the article’s Creative Commons license, unless indicated otherwise in a credit line to the
material. If material is not included in the article’s Creative Commons license and your intended use is not per-
mitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the
copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.

© The Author(s) 2019

Scientific Reports | (2019) 9:17831 | https://doi.org/10.1038/s41598-019-54386-6 12

Content courtesy of Springer Nature, terms of use apply. Rights reserved
Terms and Conditions
Springer Nature journal content, brought to you courtesy of Springer Nature Customer Service Center GmbH (“Springer Nature”).
Springer Nature supports a reasonable amount of sharing of research papers by authors, subscribers and authorised users (“Users”), for small-
scale personal, non-commercial use provided that all copyright, trade and service marks and other proprietary notices are maintained. By
accessing, sharing, receiving or otherwise using the Springer Nature journal content you agree to these terms of use (“Terms”). For these
purposes, Springer Nature considers academic use (by researchers and students) to be non-commercial.
These Terms are supplementary and will apply in addition to any applicable website terms and conditions, a relevant site licence or a personal
subscription. These Terms will prevail over any conflict or ambiguity with regards to the relevant terms, a site licence or a personal subscription
(to the extent of the conflict or ambiguity only). For Creative Commons-licensed articles, the terms of the Creative Commons license used will
apply.
We collect and use personal data to provide access to the Springer Nature journal content. We may also use these personal data internally within
ResearchGate and Springer Nature and as agreed share it, in an anonymised way, for purposes of tracking, analysis and reporting. We will not
otherwise disclose your personal data outside the ResearchGate or the Springer Nature group of companies unless we have your permission as
detailed in the Privacy Policy.
While Users may use the Springer Nature journal content for small scale, personal non-commercial use, it is important to note that Users may
not:

1. use such content for the purpose of providing other users with access on a regular or large scale basis or as a means to circumvent access
control;
2. use such content where to do so would be considered a criminal or statutory offence in any jurisdiction, or gives rise to civil liability, or is
otherwise unlawful;
3. falsely or misleadingly imply or suggest endorsement, approval , sponsorship, or association unless explicitly agreed to by Springer Nature in
writing;
4. use bots or other automated methods to access the content or redirect messages
5. override any security feature or exclusionary protocol; or
6. share the content in order to create substitute for Springer Nature products or services or a systematic database of Springer Nature journal
content.
In line with the restriction against commercial use, Springer Nature does not permit the creation of a product or service that creates revenue,
royalties, rent or income from our content or its inclusion as part of a paid for service or for other commercial gain. Springer Nature journal
content cannot be used for inter-library loans and librarians may not upload Springer Nature journal content on a large scale into their, or any
other, institutional repository.
These terms of use are reviewed regularly and may be amended at any time. Springer Nature is not obligated to publish any information or
content on this website and may remove it or features or functionality at our sole discretion, at any time with or without notice. Springer Nature
may revoke this licence to you at any time and remove access to any copies of the Springer Nature journal content which have been saved.
To the fullest extent permitted by law, Springer Nature makes no warranties, representations or guarantees to Users, either express or implied
with respect to the Springer nature journal content and all parties disclaim and waive any implied warranties or warranties imposed by law,
including merchantability or fitness for any particular purpose.
Please note that these rights do not automatically extend to content, data or other material published by Springer Nature that may be licensed
from third parties.
If you would like to use or distribute our Springer Nature journal content to a wider audience or on a regular basis or in any other manner not
expressly permitted by these Terms, please contact Springer Nature at

[email protected]